Why do I need a professional card? And that’s the right question…
I have been watching this card since the beginning of the anomalous strengthening of the ruble and managed to buy it before its expected crash! I found a store where the price in dollars corresponded to Westerners (well, almost + 15-20% normal, taking into account the logistics and fat of the store), and in rubles corresponded to the course of the Russian stock market and bought it, thanks www.regard.ru!
At first I looked at A4000 and here is a comparison with A4500 and A5000
| Model | GIGABYTE RTX 2070 MINI ITX | PNY RTX A4000 | PNY RTX A4500 | PNY RTX A5000 | RTX 3080 Ti |
|---|---|---|---|---|---|
| Chip | TU106 | GA104 | GA102 | GA102 | GA102 |
| Memory | 8Gb | 16Gb | 20Gb | 24Gb | 12Gb |
| Memory Type | GDDR6 | GDDR6 | GDDR6 | GDDR6 | GDDR6X |
| Memory Bus | 256 | 256 | 320 | 384 | 384 |
| ECC | - | + | + | + | - |
| Nvlink | - | - | + | + | - |
| vGPU | - | - | - | + | - |
| CUDA | 2304 | 6144 | 7168 | 8192 | 10240 |
| TDP | 175 W | 140W | 200W | 230W | 350 W |
| Price www.regard.ru 29.05.2021 | 46 000 RUR [Avito] | 114 310 RUR [OEM] | 148 070 RUR [OEM] | 187 780 RUR [OEM] | 165 900 RUR [RETAIL] |
A4000 seemed too slow in terms of memory, and it has a 3070 chip, without Nvlink, as a professional it only has a lot of memory and ECC.
And the A4500 chip is 3080/3090, there is Nvlink and in the future, if there is another card like it, an Nvlink adapter, a motherboard with two PCI-E, then I can make ona card from two, with 40Gb of memory. I really don’t know how it works – will it be a single card for the whole OS or just in some programs, it’s unlikely to do it anyway.
A5000 seemed too expensive for adding some CUDA and vGPU that I don’t need.
By the way, do you know why professional Nvidia video cards don’t use faster GDDR6X memory like gaming ones? Because they learned to pack only 1Gb per chip, and in the professional ones a lot of memory is needed, so here the memory is slower GDDR6, which they already know how to pack more densely! Later in the text, it will be noted that memory is everything, and speed on professional cards is not the main thing!
The card is in the OEM package, but it is not a consumer OEM in a bag and that’s it: a professional OEM is a box with pads and of course in a bag! Apparently for +22% of the price, this box is placed in a beautiful one? with a color print and is sold as RETAIL :)))
On the tactile plan: it’s a metal brick, with nice edges!
Additional power supply through a single connector
A4500 is also used for entertainment purposes, here Zotac put it in his VR gaming backpack Zotac представила компьютер-рюкзак VR GO 4.0 на базе Intel Tiger Lake и NVIDIA RTX A4500!
I will compare it with GIGABYTE RTX 2070 MINI ITX
noise and heat
I have not changed the power supply, it is the one from before Fractal Ion SFX-L 650W Gold.
I have not changed the case either, it remained the same COUGAR QBX old but very good and compact, the graphics card does not suck air from the case, but directly from the outside. And furthermore, the turbine of the A4500 immediately expels the hot air, and not like the RTX 2070 dissipates it inside. This immediately affected the rate of heating inside the box.
I was very surprised by the noise of the A4500 turbine, for its absence, the card is very quiet, compared to the RTX 2070 incredibly quiet!!! Under the heaviest load (deepfacelab), the turbine rotates only halfway.
51% 77C P2 185W / 200W
By the way, turning on additional fans like i am doing it in this article How to automatically turn on PC fans with Arduino according to temperature! goes down 3 degrees from A4500, soo it makes sense to blow the end of the turbine!
Without load, the speed of the turbine does not drop below 30%, it is quiet, but it could be less.
30% 35C P8 16W / 200W
Test of CUDA PNY RTX A4500
It has three times more CUDA cores than RTX 2070 at the same price, that is, the price of electricity, when you train a neural network for days, that matters!
The test was done on Ubuntu with geekbench5
./geekbench5 --compute CUDA
The A4500 performs better than the RTX 2070, but not three times as much (according to the number of CUDA core pieces), but almost twice as much!
Here you can see the results of other cards browser.geekbench.com/cuda-benchmarks, however, the CUDA speed of the A4500 is only 5% higher than NVIDIA GeForce RTX 3070 Ti – $898.00 at Amazon (A4500 is $1,831.61 there)… and the same as Tesla V100- FHHL- 16GB – $5,829.00. If you need CUDA speed and don’t need memory, then it’s better to buy RTX 3080 Ti…
Benchmarks of PNY RTX A4500 in games
I naively believed that it will be, although not three times as fast (because A4500 is three times more expensive than RTX 2070 and newer), but twice as fast… but no, only by 1-50%… Maybe something is wrong with the tests on older games, maybe a small screen resolution of 1920×1080 and on a larger one on RTX 2070 would be impossible play and A4500 took advantage!!!
Here are the tests, for PNY I used two drivers: the first is the one that was in the RTX 2070, the second is special for the A4500
| Games | GIGABYTE RTX 2070 MINI ITX Game Ready Driver 512.77 | PNY RTX A4500 Game Ready Driver 512.77 | PNY RTX A4500 Production Branch 512.78 |
|---|---|---|---|
| Assassin’s Creed® Odyssey [MIN] | 29 | 32 | 43 |
| Assassin’s Creed® Odyssey [MAX] | 91 | 98 | 91 |
| Assassin’s Creed® Odyssey [AVG] | 60 | 60 | 60 |
| Shadow of the Tomb Raider [MIN] | 82 | 129 | 128 |
| Shadow of the Tomb Raider [MAX] | 159 | 170 | 170 |
| Shadow of the Tomb Raider [AVG] | 106 | 154 | 153 |
| Cyberpunk 2077 [MIN] | 33 | 51 | 52 |
| Cyberpunk 2077 [MAX] | 62 | 72 | 84 |
| Cyberpunk 2077 [AVG] | 43 | 60 | 60 |
Test Screenshots – Graph at Max Default Settings
PNY RTX A4500 3dMark Score
Here at 1920 x 1080, A4500 is 55% faster than RTX 2070!
| GIGABYTE RTX 2070 MINI ITX Game Ready Driver 512.77 | PNY RTX A4500 Game Ready Driver 512.77 | PNY RTX A4500 Production Branch 512.78 | PNY RTX A4500 Production Branch 512.78 |
|
|---|---|---|---|---|
| VSync | off | off | off | on |
| 3DMark Basic Edition Graphics score | 8434 | 13102 | 13110 | 9770 |
Conclusion: In general, now I have doubts, maybe I should buy the RTX 3080 Ti it costs the same… But first, it’s too late to think about it, the price changed and also its no longer available. Second, it only has 12 Gb of memory and TDP 350W. Maybe in the long run I did the right thing by choosing one with more video memory, I’ll trade off less CUDA with processing time, which with a lower TDP, will be more profitable in terms of electricity price! By the way, if you take numbers from the geekbench5 test and divide by watts, then the A4500 will be faster per watt by about 10-15%.
…then why do I need a professional video card, which differs from gaming ones only in a larger memory capacity and loses to those(equal in price) in speed? Well just for the memory, 8Gb was not enough to increase the model in DeepFaceLab, I stopped at a resolution of 224, and now I will start feeding a new model (neural network) of 380, and its just a begin!!!
UPD: I found a way to watch PNY RTX A4500 in 4K games without a 4k monitor: via a non-4K Chinese recorder, it says 4K but only records in 1080p. It is clear that the gaming performance of the PNY RTX A4500 is between the RTX 3070 and the RTX 3060 TI, but still… I even made the video through this same recorder!
ECC is disabled, Vsync is also disabled, which is why the image is sometimes broken, terrible, but to measure pure performance, it’s supposed to be off!
3Dmark da una cifra más o menos la misma, pensé que depende de la resolución de la pantalla, pero no es así
And in games, of course, everything has changed!
Cyberpunk 2077 less 55%
Shadow of the Tomb Raider less 45%
Assassin’s Creed® Odyssey less 18%
| 4K ULtra Hight DLSS Off | MIN | AVG | MAX |
|---|---|---|---|
| Cyberpunk 2077 | 18 | 27 | 37 |
| Shadow of the Tomb Raider | 48 | 59 | 85 |
| Assassin’s Creed® Odyssey | 31 | 49 | 77 |
Leave a Reply